Selective Transfer Learning for Cross Domain Recommendation
Abstract
Collaborative filtering (CF) aims to predict users’ ratings on items according to historical useritem preference data. In many realworld applications, preference data are usually sparse, which would make models overfit and fail to give accurate predictions. Recently, several research works show that by transferring knowledge from some manually selected source domains, the data sparseness problem could be mitigated. However for most cases, parts of source domain data are not consistent with the observations in the target domain, which may misguide the target domain model building. In this paper, we propose a novel criterion based on empirical prediction error and its variance to better capture the consistency across domains in CF settings. Consequently, we embed this criterion into a boosting framework to perform selective knowledge transfer. Comparing to several stateoftheart methods, we show that our proposed selective transfer learning framework can significantly improve the accuracy of rating prediction tasks on several realworld recommendation tasks.
Selective Transfer Learning for Cross Domain Recommendation
Zhongqi Lu^{†}^{†}thanks: Hong Kong University of Science & Technology. {zluab, ezhong, skyezhao, wxiang, weikep, qyang}@cse.ust.hk and Erheng Zhong^{†}^{†}footnotemark: and Lili Zhao^{†}^{†}footnotemark: and Evan Xiang^{†}^{†}footnotemark: and Weike Pan^{†}^{†}footnotemark: and Qiang Yang^{†}^{†}footnotemark: 
Keywords: Transfer Learning; Collaborative Filtering; Cross Domain Recommendation;
1 Introduction
Recommendation systems attempt to recommend items (e.g., movies, TV, books, news, images, web pages, etc.) that are likely to be of interest to users. As a stateoftheart technique for recommendation systems, collaborative filtering aims at predicting users’ ratings on a set of items based on a collection of historical useritem preference records. In the realworld recommendation systems, although the item space is often very large, users usually rate only a small number of items. Thus, the available rating data can be extremely sparse for each user, which is especially true for new online services. Such data sparsity problem may make CF models overfit the limited observations and result in lowquality predictions.
In recent years, different transfer learning techniques have been developed to improve the performance of learning a model via reusing some information from other relevant systems for collaborative filtering [11, 27]. And with the increasing understandings of auxiliary data sources, some works (like [5, 22]) start to explore data from multiple source domains to achieve more comprehensive knowledge transfer. However, these previous methods overtrust the source data and assume that the source domains follow the very similar distributions with the target domain, which is usually not true in the realworld applications, especially under the cross domain CF settings. For example, in a local music rating web site, natives may give trustful ratings for the traditional music; while in an international music rating web site, the ratings on those traditional music could be diverse due to the culture differences: those users with good culture background would constantly give trustful ratings, others could be inaccurate. If the target domain task is the music recommendation of a startup local web site, obviously we do not want all the International web site’s data as source domain without selection. To better tackle the cross domain CF problems, we face the challenge to tell how consistent the data of target and source domains are and adopt only those consistent source domain data while transferring knowledge.
Several research works (like [2]) have been proposed to perform instance selection across domains for classification tasks based on empirical error. But they cannot be adopted to solve CF problems directly. Especially when the target domain is sparse, because of the limited observations of user’s ratings on the items in the target domain, getting a low empirical error occasionally in the target domain does not mean the source domains are truly helpful in building a good model. In other words, the inconsistent knowledge from source domains may dominate the target domain model building and happen to fit the few observations in the target domain, which gives high accuracy unacceptably.
We take careful analysis on this problem and in our observation on the music rating example, some users, such as domain experts, follow standard criteria to rate and hence share a consistent distribution over the mutual item set across domains. And further, we find this consistency can be better described by adding the variance of empirical error produced by the model. The smaller the variance of empirical error on predictions for a user, the more likely this user is consistent with those from other domains. And we would like to adopt those who are more likely to share consistent preferences to perform knowledge transfer across domains. Based on this observation, we propose a new criterion using both empirical error and its variance to capture the consistency between source and target domains. As an implementation, we embed this criterion into a boosting framework and propose a novel selective transfer learning approach for collaborative filtering (STLCF). STLCF works in an iterative way to adjust the importance of source instances, where those source data with low empirical error as well as low variance will be selected to help build target models.
Our main contributions are summarized as follows:

First, we find that selecting consistent auxiliary data for the target domain is important for the crossdomain collaborative filtering, while the consistency between source and target domain is influenced by multiple factors. To describe these factors, we propose a novel criterion, based on both empirical error and its variance.

Second, we propose a selective transfer learning framework for collaborative filtering  an extension of the boosting based transfer learning algorithm that take the above factors into consideration, so that the sparseness issue in the CF problems can be better tackled.

Third, the proposed framework is general, where different base models can be embedded. We propose an implementation based on Gaussian probability latent semantic analysis, which demonstrates the proposed framework can solve the sparseness problem on various realworld applications.
2 Preliminaries
2.1 Problem Settings
Suppose that we have a target task where we wish to solve the rating prediction problem. Taking the regular recommendation system for illustration, is associated with users and items denoted by and respectively. In this task, we observe a sparse matrix with entries . Let denote the set of observed links in the system. For the rating recommendation system, can either take numerical values, for example , or binary values . We aim to transfer knowledge from other source domains with each source domain contains users and items denoted by and . Similar to the target domain, each source domain contain sparse matrices and observed links .
The settings of STLCF are illustrated in Figure 1. We adopt a setting commonly used in transfer learning for collaborative filtering: either the items or the users that are in the target domain also appear in the source domains. In the following derivation and description of our STLCF model, for the convenience of interpretation, we focus on the case that the user set is shared by both target domain and the source domains. The case that the item set is shared can be easily tackled in a similar manner.
Under the assumption that the observation is obtained with and being independent, we formally define a cooccurrence model in both the target and the source domains to solve the collaborative filtering problem:
In the following, based on Gaussian probabilistic latent semantic analysis (GPLSA), we first briefly present a transfer learning model for collaborative filtering  transferred Gaussian probabilistic latent semantic analysis (TGPLSA) as an example, which is designed to integrate into our later proposed framework as a base model. After that, we present our selective transfer learning for collaborative filtering (STLCF) to perform knowledge transfer by analyzing the inconsistency between the observed data in target domain and the source domains. Careful readers shall notice that other than the TGPLSA example, STLCF is compatible to use various generative models as the base model.
2.2 Collaborative Filtering via Gaussian Probabilistic Latent Semantic Analysis (GPLSA)
For every useritem pair, we introduce hidden variables with latent topics , so that user and item are rendered conditionally independent. With observations of item set , user set and rating set in the source domain, we define a model as:
We further investigate the use of a Gaussian model for estimating by introducing for the mean and for the variance of the ratings. With these, we define a Gaussian mixture model for a single domain as:
where is the topic distribution over users, and follows a Gaussian distribution.
Maximum likelihood estimation amounts to minimize:
(2.1) 
where is a generic parameter referring to a particular model.
Next, we extend GPLSA to the cross domain context to achieve transfer learning for collaborative filtering (TLCF).
2.3 Transfer Learning for Collaborative Filtering (TLCF)
When the target data is sparse, GPLSA may overfit the limited observed data. Following the similar idea in [25], we extend GPLSA to the Transferred Gaussian Probabilistic Latent Semantic Analysis (TGPLSA) model. Again we use to denote index of the source domain where the knowledge come from, and to denote the index of the target domain where the knowledge is received. For simplicity, we present the work with one source domain, and this model can be easily extended to multiple source domains. Moreover, we assume all the users appear in both the source domains and the target domain. Such scenarios are common in the realworld systems, like Douban^{1}^{1}1http://www.douban.com  a widely used online service in China, which provides music, book and movie recommendations..
Similar to the approach in [25], the TGPLSA jointly learn the two models for both the source domain and the target domain using a relative weight parameter^{2}^{2}2, which is introduced to represent the tradeoff of source and target information. . Since the item sets and are different or even disjoint with each other, there could be inconsistency across domains as we discussed in Section 1. Clearly, the more consistent source and target domains are, the more help target task could get from source domain(s). We are motivated to further analyze this inconsistency in our work by learning item weight vectors and of the instances in source and target domain respectively. Then, the objective function in Eq.(2.1) can be extended as:
(2.2)  
We adopt the expectationmaximization (EM) algorithm, a standard method for statistical inference, to find the maximum loglikelihood estimates of Eq.(2.2). Details of derivations can be found in the appendix.
3 Selective TLCF
As we have discussed before, using the source domain data without selection may harm the target domain learning. By proposing the selective knowledge transfer with the novel factors (empirical error and variance of empirical error), we come up with the details of Selective Transfer Learning framework for CF in this section. As illustrated in the second example in Figure 1 where the domains have mutual user set, we would like to transfer knowledge of those items’ records that consistently reflect the user’s preferences. Because of our finding that the consistent records have small empirical error and variance, the selection shall consider these two factors. We embed these two factors into a boosting framework, where the source data with small empirical error and variance receive higher weights since they are consistent with the target data. This boosting framework models the crossdomain CF from two aspects: on one hand, we take more care of those mispredicted target instances; on the other hand, we automatically identify the consistency of the source domains during the learning process and selective use those source domains with more trustful information.
As shown in Algorithm 1, in each iteration, we apply base model TGPLSA over weighted instances to build a weak learner and hypothesis . Then to update the source and target item weights, domain level fitness weight is chosen for each source domain based on domain level consistency [4]. And for base model is also updated, considering empirical errors and variances. Accordingly, the weights of mispredicted target items are increased and the weights of those less helpful source domains are decreased in each iteration. The final ensemble is given by an additive model, which gives larger weights to the hypotheses with lower errors. We provide a detailed derivation of STLCF in the rest of this section.
In previous works in collaborative filtering, the mean absolute error (MAE) is usually chosen as the loss function. If we tolerate some prediction error , we define:
(3.3) 
where is the number of observed ratings. and denote the true values and predictions respectively. We may also define the item level MAE error for target domain with respect to as:
(3.4) 
To facilitate the optimization, we consider the following exponential loss for empirical risk minimization:
(3.5) 
As stated in previous section, the lower variance of empirical errors can provide more confident consistency estimation, we combine these factors and reformulate the loss function:
(3.6) 
Above all, the model minimize the above quantity for some scalar :
Assume that the function of interest for prediction is composed of the hypothesis from each weak learner. The function to be output would consist of the following additive model over the hypothesis from the weak learners:
(3.7) 
where .
Since we are interested in building an additive model, we assume that we already have a function . Subsequently, we derive a greedy algorithm to obtain a weak learner and a positive scalar such that .
In the following derivation, for the convince of presentation, we omit the model index , and use to represent , to represent .
By defining , , , and , Eq.(3.6) can be equivalently posed as optimizing the following loss with respect to :
(3.8)  
For brevity, we define the following sets of indices as and . Here denotes the set of items whose prediction by falls into the fault tolerable range, while denotes the rest set. By making the last transformation in Eq.(3.8) equal to zero, we get:
(3.9)  
If we set , then it is reduced to the form of AdaBoost:
(3.10)  
Finally, the updating rule for is
(3.11) 
And for the instance weight in the source domain, we can also adopt the similar updating rule in Eq.(3.11).
Other than the instance level selection discussed above, we also want to perform the domain level selection to penalize those domains that are likely to be irrelevant, so that the domains with more relevant instances speak loudly. Following the idea of taskbased boosting [3], we further introduce a reweighting factor for each source domain to control the knowledge transfer. So we formulate the updating rule for to be:
(3.12) 
where can be set greedily in proportion to the performance gain of the single source domain transfer learning:
(3.13) 
where is the training error of the transfer learning model, and is the training error of the nontransfer learning model, which utilizes only the observed target domain data.
4 Experiments
Notation  Data Set  Data Type  Instances No. 

D1  Douban Music  Rating [1,5]  
D2  Douban Book  Rating [1,5]  
D3  Douban Movie  Rating [1,5]  
D4  Netflix  Rating [1,5]  
D5  Wikipedia  Editing Log  
D6  IMDB  Hyperlink 
4.1 Data Sets and Experimental Settings
We evaluate the proposed method on four data sources: Netflix^{3}^{3}3http://www.netflix.com, Douban IMDB^{4}^{4}4http://www.imdb.com, and Wikipedia^{5}^{5}5http://en.wikipedia.org user editing records. The Netflix rating data contains more than million ratings with values in , which are given by more than users on around movies. Douban contains movie, book and music recommendations, with rating values also in . IMDB hyperlink graph is employed as a measure of similarity between movies. In the graph, each movie builds links to its most similar movies. The Wikipedia user editing records provide a indicator of whether a user concerns or not about a certain movie.
The data sets used in the experiments are described as follows. For Netflix, to retain the original features of the users while keeping the size of the data set suitable for the experiments, we sampled a subset of users. In Douban data sets, we obtained ratings on music, ratings on books, and ratings on movies, given by users. For both the IMDB data set and the Wikipedia data set, we filtered them by matching the movie titles in both the Netflix and the Douban Movie data sets. After preprocessing, the IMDB hyperlink data set contains movies. The Wikipedia user editing records data set has editing logs by users on the same movies as IMDB data set. To present our experiments, we use the shorthand notations listed in Table 1 to denote the data sets.
We evaluate the proposed algorithm on five crossdomain recommendation tasks, as follows:

The first task is to simulate the crossdomain collaborative filtering, using the Netflix data set. The sampled data is partitioned into two parts with disjoint sets of movies but identical set of users. One part consists of ratings given by movies with density, which serves as the source domain. The remaining movies are used as the target domain with different levels of sparsity density.

The second task is a realworld crossdomain recommendation, where the source domain is Douban Book and the target domain is Douban Movie. In this setting, the source and the target domains share the same user set but have different item sets.

The third task is constructed with Netflix and Douban data. There are about shared movies in Netflix and Douban Movie. We extract the ratings on shared movies from Netflix and Douban Movie. Then we get ratings from Douban given by users with density , and ratings from Netflix given by users with density . The goal is to transfer knowledge from the Netflix data set to Douban Movie. In this task, item set is identical across domains but user sets are totally different.

The fourth task is to evaluate the effectiveness of the proposed algorithm under the context of multiple source domains. It uses both Douban Music and Douban Book as the source domains and transfer knowledge to Douban Movie domain.

The fifth task varies the type of source domains. It utilizes the Wikipedia user editing records and IMDB hyperlink graph, together with Douban Movie as the source domains to perform rating predictions on the Netflix movie data set.
For evaluation, we calculate the Root Mean Square Error (RMSE) on the heldout of the target data:
where and are the true and predicted ratings, respectively, and is the number of test ratings.
Datasets 
Source  Target  NonTL  NonSelective TL  Selective TL  
sparseness  sparseness  GPLSA  PMF  TGPLSA  CMF  STLCF(E)  STLCF(EV)  
D4(Simulated)  1.6%  0.1%  1.0012  0.9993  0.9652  0.9688  0.9596  0.9533 
to  0.2%  0.9839  0.9814  0.9528  0.9532  0.9468  0.9347  
D4(Simulated)  0.3%  0.9769  0.9728  0.9475  0.9464  0.9306  0.9213  
1.5%  0.1%  0.8939  0.8856  0.8098  0.8329  0.7711  0.7568  
D2 to D3  0.2%  0.8370  0.8323  0.7462  0.7853  0.7353  0.7150  
0.3%  0.7314  0.7267  0.7004  0.7179  0.6978  0.6859  
1.7%  0.1%  0.8939  0.8856  0.8145  0.8297  0.7623  0.7549  
D4 to D3  0.2%  0.8370  0.8323  0.7519  0.7588  0.7307  0.7193  
0.3%  0.7314  0.7267  0.7127  0.7259  0.6982  0.6870  

4.2 STLCF and Baselines Methods
We implement two variations of our STLCF method. STLCF(E) is an STLCF method that only take training error into consideration when performing selective transfer learning. STLCF(EV) not only considers training error, but also utilizes the empirical error variance. To demonstrate the significance of our STLCF, we selected the following baselines^{6}^{6}6Parameters for these baseline models are finetuned via cross validation.: PMF [20] is a recently proposed method for missing value prediction. Previous work showed that this method worked well on the large, sparse and imbalanced data set. GPLSA [7] is a classical nontransfer recommendation algorithm. CMF [24] is proposed for jointly factorizing two matrices. Being adopted as a transfer learning technique in several recent works, CMF has been proven to be an effective crossdomain recommendation approach. TGPLSA is an uniformly weighted transfer learning model, which utilize all source data to help build the target domain model. It is used as one of the baselines because we adopt it as the base model of our boostingbased selective transfer learning framework.
4.3 Experimental Results
4.3.1 Performance Comparisons
We test the performance of our STLCF methods against the baselines. The results of the collaborative filtering tasks under three different target domain sparseness are shown in Table 2.
First, we observe that the nontransfer methods, i.e. GPLSA and PMF, fail to give accurate predictions, especially when the target domain is severely sparse. With the help of source domains, the (nonselective) transfer learning methods with equally weights on the source domains, like TGPLSA and CMF, can increase the accuracy of the rating predictions. And our selective transfer learning methods (i.e., STLCF(E) and STLCF(EV)) can do even better. The fact that our STLCF outperforms others is expected because by performing the selective knowledge transfer, we use the truly helpful source domain(s), which is designed to handle the sparseness issue in CF problems.
Second, comparing the two nonselective TLCF methods with the other two selective TLCF methods, we observe that on the last two real world tasks (D2 to D3 and D4 to D3) when the target domain is extremely sparse (say 0.1%), the improvement of accuracy achieved by our STLCF methods against the nonselective transfer learning methods is much more significant than it does on the simulation data set based on Netflix (D4 to D4). Notice that the inconsistency of the target domain and the source domains on the simulation data sets is much smaller than that on the realworld cases. The experiment results show that our STLCF algorithm is effective in handling the inconsistency between the sparse target domain and the source domains.
Third, we notice that some factors, like empirical error variance, may affect the prediction. In Table 2, we compare our two STLCF methods, i.e., STLCF(E) and STLCF(EV) when the target domain sparsity is . We can find that on the task “D2 to D3”, i.e., Douban Book to Movie, STLCF(EV) is much better than STLCF(E). But on the task “D4(Simulated) to D4(Simulated)”, the improvement of STLCF(EV) is not so significant against STLCF(E). These observations may be due to the domain consistency. For the tasks “D4(Simulated) to D4(Simulated)”, both the source and target entities are movie ratings from Netflix data set, while the task “D2 to D3” tries to transfer the knowledge from a book recommendation system to the moive recommendation system, which may contain some domain specific items. When the target domain is very sparse, i.e. the user’s ratings on the items are rare, there are chances to get high prediction accuracy occasionally on the observed data with a bad model on the source domains that are inconsistent with target domain. In this case, it is important to consider the variance of empirical error as well. Comparing to STLCF(E), STLCF(EV), which punishes the large variance, can better handle the domain inconsistency in transfer learning, especially when the target domain is sparse.
4.3.2 Results on LongTail Users
Ratings  NonTL  NonSelective TL  Selective TL  

per  i.e. STLCF  
user  GPLSA  TGPLSA  CMF  (E)  (EV) 
15  1.1942  0.9294  0.9312  0.8307  0.8216 
610  0.9300  0.7859  0.7929  0.7454  0.7428 
1115  0.8296  0.7331  0.7390  0.7143  0.7150 
1620  0.7841  0.7079  0.7113  0.7042  0.7050 
2125  0.7618  0.6941  0.6947  0.6942  0.6910 
2630  0.7494  0.6918  0.6884  0.6917  0.6852 
3135  0.7370  0.6909  0.6911  0.6915  0.6818 
3640  0.7281  0.6896  0.6856  0.6907  0.6776 
4145  0.7219  0.6878  0.6821  0.6890  0.6740 
4650  0.7187  0.6881  0.6878  0.6800  0.6734 
To better understand the impact of STLCF with the help of the source domain, we conduct a finegrained analysis on the performance improvement on Douban data sets, with Douban Book as source domain and Douban Movie as target domain. The results on different user groups in the target domain are shown in Table 3. First, we observe that the STLCF models, i.e., STLCF(E) and STLCF(EV) can achieve significantly better results on those longtail users who have very few ratings in historical logs. Such fact implies that our STLCF methods could handle the longtail users that really need a finegrained analysis when performing knowledge transfer from source domains. Current CF models without any finegrained analysis on the specific users usually fail to capture the preferences of the longtail users, while our STLCF methods work well because they can selectively augment the weight of the corresponding source domain instances with respect to those longtail cases at both instance level and domain level. Second, STLCF(EV) works better than STLCF(E) on those nonlongtail users, i.e., with more than 25 ratings per user in the historical log. This is expected because users with more ratings can benefit more from the error variance analysis to avoid negative knowledge transfer.
Source Domain:  None  D3  D3 & D5  D3 & D6  D5 & D6  D3 & D5 & D6  

Target  0.1%  0.9983  0.9663  
(D4)  0.2%  0.9812  0.9505  
sparseness  0.3%  0.9703  0.9383 
Source Domain:  None  D1  D2  D1 & D2  

Target  0.1%  0.8856  0.7304  
(D3)  0.2%  0.8323  0.6904  
sparseness  0.3%  0.7267  0.6739 
4.3.3 STLCF with Multiple Source Domains
We apply STLCF(EV) on the extremely sparse target movie domain, with two sets of source domains: one is composed of Douban Music and Douban Book, the other is composed of Douban Movie, IMDB hyperlink graph and Wikipedia user editing records. The results are in Table 5 and Table 4 respectively. We demonstrate our STLCF method can utilize multiple source domains of various types by handling the inconsistency between the target and the source domains.
First, for the Douban experiments shown in Table 5, we observe that comparing to only using either Douban Book or Douban Music as source domain, there are significant improvements when both of them are used. The result is expected because each of the source domains has its own parts of effective information for the target domain. For example, a user who show much interests in the movie “The Lord of the Rings” may have consistent preferences in its novel. In this case, with the help of more auxiliary sources, better results are expected.
Second, we explore the generalization of the choices of source domains by introducing domains like Wikipedia user editing records and IMDB hyperlink graph, which are not directly related to the target domain but still contain some useful information in helping the target task (Netflix rating prediction). The results are shown in Table 4. Comparing the results of the experiment that uses no source domain (nontransfer) to those that use source domains D5 & D6, we observe that although the Wikipedia user editing records or IMDB hyperlink graph is not closely related to the target domain and can hardly be adopted as source domains by previous transfer learning techniques, our STLCF method can still transfer useful knowledge successfully. In addition, comparing the results of the experiment that uses single source domain D3 to those that use source domains D3 & D5, D3 & D6, or D3 & D5 & D6, we find that the Wikipedia user editing records or IMDB hyperlink graph could provide some useful knowledge that is not covered by the related movie source domains. Despite of the noise and heterogeneous setting, our STLCF method can still utilize these source domains to help the target domain tasks. As we have discussed in Section 3, our STLCF performs selective transfer learning at both domain level and instance level. On one hand, the domain level selective transfer can block the noisy information globally. As we can see, D5 & D6 are noisy and therefore contain much data that are inconsistent with the observed data in the target domain, therefore the overall transfer of D5 & D6 is penalized. On the other hand, the instance level selective transfer learning can further eliminate the affections of those irrelevant source instances.
Above all, our STLCF is highly adaptive to utilize source domains that are relatively inconsistent with the target domain, even when the target domain is rather sparse.
4.3.4 Parameters Analysis of STLCF
There are two parameters in our STLCF, i.e., the prediction error threshold and the empirical error variance weight . Since and are independent, we fix one and adjust another.
We fix the empirical error variance weight to be and adjust the parameter . Based on our results shown in Figure 2, the model has good performance when is of order . We also tuned the parameter , which balances the empirical error and its variance. We fix the prediction error threshold to be in tuning . As shown in Figure 3, when we vary the parameter from to , the best choices of are found to be around .
4.3.5 Convergence and Overfitting Test
Figure 4 shows the RMSEs of STLCF(EV) as the number of weak learners changes on the Douban Book to Movie task. From the figure on the left, we observe that STLCF(EV) converges well after 40 iterations. We can also find that the corresponding also converge to around 0.68 after 40 iterations as well.
The number of latent topics of the base learner TGPLSA reflects the model’s ability to fit training data. When we keep increasing the number of latent topics, the model tends to better fit the training data. But if the number of latent topics is too large, the model may suffer from overfitting. We investigate the overfitting issue by plotting the training and testing RMSEs of the nontransfer learning model GPLSA, the nonselective transfer learning model TGPLSA and our selective transfer learning model STLCF(EV) over different numbers of latent topics in Figure 5. The data sparsity for the target domain is around 0.3%.
We can observe that comparing to our STLCF, the training RMSEs of GPLSA and TGPLSA decrease faster, while their testing RMSEs go down slower. When is about , the testing RMSEs of GPLSA start to go up. And for TGPLSA, its testing RMSEs also go up slightly when is larger than . But the testing RMSEs of our STLCF keep decreasing until and even when is larger than , the raise of our STLCF’s testing RMSEs is not obvious. Clearly when the target domain is very sparse, our STLCF method is more robust against the overfitting, by inheriting the advantage from boosting techniques and the finegrained selection on knowledge transfer.
5 Related Works
Selective  NonSelective  

Transfer  STLCF  RMGM [11], CMF [24], 
Learning  TIF [15], etc.  
NonTransfer  –  MMMF [19], GPLSA [7], 
Learning  PMF [20], etc. 
The proposed Selective Transfer Learning for Collaborative Filtering (STLCF) algorithm is most related to the works in collaborative filtering. In Table 6, we summarize the related works under the collaborative filtering context. To the best of our knowledge, no previous work for transfer learning on collaborative filtering has ever focused on the finegrained analysis of consistency between source domains and the target domain, i.e., the selective transfer learning.
Collaborative Filtering as an intelligent component in recommender systems has gained extensive interest in both academia and industry. Various models have been proposed, including factorization models [10, 15, 16, 18, 19], probabilistic mixture models [8, 9], Bayesian networks [17] and restricted Boltzman machines [21]. However, most of the previous work would suffer from overfitting to the small set of observed data. In this paper, we introduce the concept of selective transfer learning to better tackle the overfitting and data sparseness issue.
Transfer Learning [14] utilizes data sets from related but different domains to build model for a target application domain. A few works on transfer learning are in the context of collaborative filtering. Mehta and Hofmann [13] consider the scenario involving two systems with shared users and use manifold alignment methods to jointly build neighborhood models for the two systems. They focus on making use of an auxiliary recommender system when only part of the users are aligned, which does not distinguish the consistency of users’ preferences among the aligned users. Li et al. [12] designed a regularization framework to transfer knowledge of clusterlevel rating patterns, which does not make use of the correspondence between source and target domains.
Recently, researchers propose the MultiSourceTrAdaBoost [26] to allow automatically selecting the appropriate data for knowledge transfer from multiple sources. The newest work TransferBoost [3] was proposed to iteratively construct an ensemble of classifiers via reweighting source and target instance via both individual and taskbased boosting. Moreover, EBBoost [23] suggests weight the instance based on the empirical error as well as its variance. However so far, the works limit to the classification tasks. Our work is the first to systematically study selective knowledge transfer in the settings of collaborative filtering. Besides, we propose the novel factor  variance empirical error that is proven to be of much help in solving the real world CF problems.
6 Conclusions
In this paper, we proposed to perform selective knowledge transfer for CF problems and came up with a systematical study on how the factors such as variance of empirical error could leverage the selection. We found although empirical error is effective to model the consistency across domains, it would suffer from the sparseness problem in CF settings. By introducing a novel factor  variance of empirical error to measure how trustful this consistency is, the proposed criterion can better identify the useful source domains and the helpful proportions of each source domain. We embedded this criterion into a boosting framework to transfer the most useful information from the source domains to the target domain. The experimental results on realworld data sets showed that our selective transfer learning solution performs significantly better than several stateoftheart methods at various sparsity levels. Furthermore, comparing to existing methods, our solution works well on longtail users and is more robust to overfitting.
References
 [1] W. Burgard and D. Roth, editors. Proceedings of the TwentyFifth AAAI Conference on Artificial Intelligence, AAAI 2011, San Francisco, California, USA, August 711, 2011. AAAI Press, 2011.
 [2] W. Dai, Q. Yang, G.R. Xue, and Y. Yu. Boosting for transfer learning. In Ghahramani [6], pages 193–200.
 [3] E. Eaton and M. desJardins. Selective transfer between learning tasks using taskbased boosting. In Burgard and Roth [1].
 [4] E. Eaton, M. desJardins, and T. Lane. Modeling transfer relationships between learning tasks for improved inductive transfer. In The European Conference on Machine Learning, pages 317–332, 2008.
 [5] H. Eldardiry and J. Neville. Acrossmodel collective ensemble classification. In Burgard and Roth [1].
 [6] Z. Ghahramani, editor. Machine Learning, Proceedings of the TwentyFourth International Conference (ICML 2007), Oregon, USA, June 2024, 2007, volume 227 of ACM International Conference Proceeding Series. ACM, 2007.
 [7] T. Hofmann. Collaborative filtering via gaussian probabilistic latent semantic analysis. In SIGIR, pages 259–266, 2003.
 [8] T. Hofmann. Latent semantic models for collaborative filtering. ACM Trans. Inf. Syst., 22(1):89–115, 2004.
 [9] R. Jin, L. Si, C. Zhai, and J. Callan. Collaborative filtering with decoupled models for preferences and ratings. In Proceedings of CIKM 2003, 2003.
 [10] Y. Koren, R. Bell, and C. Volinsky. Matrix factorization techniques for recommender systems. Computer, 42(8):30–37, 2009.
 [11] B. Li, Q. Yang, and X. Xue. Can movies and books collaborate?: crossdomain collaborative filtering for sparsity reduction. In Proceedings of the 21st international jont conference on Artifical intelligence, pages 2052–2057, 2009.
 [12] B. Li, Q. Yang, and X. Xue. Transfer learning for collaborative filtering via a ratingmatrix generative model. In International Conference on Machine Learning, volume 26, 2009.
 [13] B. Mehta and T. Hofmann. Cross system personalization and collaborative filtering by learning manifold alignments. In KI 2006: Advances in Artificial Intelligence, pages 244–259. 2007.
 [14] S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345–1359, October 2010.
 [15] W. Pan, E. W. Xiang, and Q. Yang. Transfer learning in collaborative filtering with uncertain ratings. In AAAI, 2012.
 [16] A. Paterek. Improving regularized singular value decomposition for collaborative filtering. Proceedings of KDD Cup and Workshop, 2007.
 [17] D. M. Pennock, E. Horvitz, S. Lawrence, and C. L. Giles. Collaborative filtering by personality diagnosis: A hybrid memory and modelbased approach. In Proc. of UAI, pages 473–480, 2000.
 [18] S. Rendle. Factorization machines with libfm. ACM Transactions on Intelligent Systems and Technology (ACM TIST), 3:19:1–19:22, May 2012.
 [19] J. D. M. Rennie and N. Srebro. Fast maximum margin matrix factorization for collaborative prediction. In ICML, pages 713–719, 2005.
 [20] R. Salakhutdinov and A. Mnih. Probabilistic matrix factorization. In NIPS, 2007.
 [21] R. Salakhutdinov, A. Mnih, and G. E. Hinton. Restricted boltzmann machines for collaborative filtering. In Ghahramani [6], pages 791–798.
 [22] X. Shi, J.F. Paiement, D. Grangier, and P. S. Yu. Learning from heterogeneous sources via gradient boosting consensus. In SDM, pages 224–235. SIAM / Omnipress, 2012.
 [23] P. K. Shivaswamy and T. Jebara. Empirical bernstein boosting. Journal of Machine Learning Research  Proceedings Track, 9:733–740, 2010.
 [24] A. P. Singh and G. J. Gordon. Relational learning via collective matrix factorization. In KDD, pages 650–658, 2008.
 [25] G.R. Xue, W. Dai, Q. Yang, and Y. Yu. Topicbridged plsa for crossdomain text classification. In S.H. Myaeng, D. W. Oard, F. Sebastiani, T.S. Chua, and M.K. Leong, editors, SIGIR, pages 627–634. ACM, 2008.
 [26] Y. Yao and G. Doretto. Boosting for transfer learning with multiple sources. In CVPR, pages 1855–1862, 2010.
 [27] Y. Zhang, B. Cao, and D.Y. Yeung. Multidomain collaborative filtering. In UAI, pages 725–732, 2010.